Báo cáo hóa học: " Research Article View Influence Analysis and Optimization for Multiview Face Recognition" pptx

Hindawi Publishing Corporation
EURASIP Journal on Image and Video Processing
Volume 2007, Article ID 25409, 8 pages
doi:10.1155/2007/25409
Research Article
View Influence Analysis and Optimization for
Multiview Face Recognition
Won-Sook Lee
1
and Kyung-Ah Sohn
2
1
School of Information Technology and Engineering, University of Ottawa, Ottawa, Canada K1N6N5
2
Computer Science Department, Carnegie Mellon University, Pittsburgh, PA 15213-3891, USA
Received 1 May 2006; Revised 20 December 2006; Accepted 24 June 2007
Recommended by Christophe Garcia
We present a novel method to recognize a multiview face (i.e., to recognize a face under different views) through optimization of
multiple single-view face recognitions. Many current face descriptors show quite satisfactory results to recognize identity of people
with given limited view (especially for the frontal view), but the full view of the human head has not yet been recognizable with
commercially acceptable accuracy. As there are various single-view recognition techniques already developed for very high success
rate, for instance, MPEG-7 advanced face recognizer, we propose a new paradigm to facilitate multiview face recognition, not
through a multiview face recognizer, but through multiple single-view recognizers. To retrieve faces in any view from a registered
descriptor, we need to give corresponding view information to the descriptor. As the descriptor needs to provide any requested
view in 3D space, we refer to it as “3D” information that it needs to contain. Our analysis in various angled views checks the extent
of each view influence and it provides a way to recognize a face through optimized integ ration of single view descriptors covering
the view plane of horizontal rotation from
−90
◦
to 90
◦

and vertical rotation from −30
◦
to 30
◦
. The resulting face descriptor based
on multiple representative views, which is of compact size, shows reasonable face recognition performance on any view. Hence,
our face descriptor contains quite enough 3D information of a person’s face to help for recognition and eventually for search,
retrieval, and browsing of photographs, videos, and 3D-facial model databases.
Copyright © 2007 W S. Lee and K A. Sohn. This is an open access article distributed under the Creative Commons Attribution
License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly
cited.
1. INTRODUCTION
Face recognition techniques have started to be used as com-
mercial products in the last few years, especially on the
frontal images, but with certain constr aints such as indoor
environment, controlled illumination, and small degree of
facial expression as can be seen in many literatures, for ex-
ample, in a classic survey paper by Samal and Iyengar [1].
Face recognition is composed of two main steps, registration
and retrieval. We register a person’s face in a certain form,
and we retrieve the person’s face out of many people’s faces.
One problem we want to raise in this paper is what is the op-
timized way to determine how many views and which angle
we need to register the person to retrieve the person in any
angle. As an effort to make more practical systems, various
researches have been performed to detect and recognize faces
in arbitrary poses or views. However, those approaches using
statistical learning methods [2–4] reveal limitation to satisfy
practically acceptable recognition performance. Novel view
generation using 3D morphable model approach [5] shows

quite reasonable success rate in many different views, but it
still depends on the database of 3D generic models to build
the linear interpolation of a given person and also it needs
high computational costs with very complicated algorithms
behind. Recently, 3D face model from direct 3D scanning
could be used for face recognition [6–8], but the successful
reconstruction is not always guaranteed in real time and the
recognition rate is not yet as good as the 2D-image-based
face recognition. In addition, the acquisition of the data is
notalwaysaseasyasimagesandwestillneedmorerobust
and stable sensing equipment to get meaningful recognition
applications. In short, multiview face recognition has still a
lot lower recognition rate compared to single view recogni-
tion.
As a representative method of the currently available
2D-based face descriptor, MPEG-7 advanced f ace recognizer
[9, 10] shows quite satisfactory results to recognize identity
of people with given single view, and it especially shows good
performance on the frontal view. However, the single-view-
based face descriptor, as it allows only one view to build its
2 EURASIP Journal on Image and Video Processing
2
−1
−0.5
0.5
1
x
−1
−0.5
0

0.5
1
z
−1
−0.5
0
0.5
1
y
Region of quasi-0
◦
horizontal view
Region of quasi-30
◦
horizontal view
View region recognized by the given image
v
Figure 1: Single view recognition of the view-sphere sur face.
(0.3, 0.32)
(0.7, 0.32)
−30
◦
−20
◦
−10
◦
0
◦
10
◦

20
◦
30
◦
−90
◦
0
◦
90
◦
Figure 2: Eye positions on view mosaic of faces from 108 ren-
dered images of 3D facial mesh models. Left eye position keeps
(0.7, 0.32) for positive horizontal rotation while right eye position
does (0.3, 0.32) for negative hor izontal rotation when width and
height of the image are considered as 1.0.
descriptor, causes problems to recognize other views. Never-
theless, it still allows nearby frontal views recognizable with
desirable success rate.
In this paper, we present a novel face descriptor based
on multiple single-view recognition, which a ims to contain
multiview 3D information of a person to help for face recog-
nition in any view. In this scenario, we save or register the face
descriptor as unique information of each person, and when
we have a query face image in arbitrary view, we can identify
the person’s identity by comparing the registered descriptors
with the one extracted from the query image. To retrieve such
3D information of a face to be recognizable in any view, we
propose a method to extend the traditional 2D-image-based
face recognition to 3D by combining multiple single views.
We take a systematic extension to build 3D information us-

ing multiviews and perform optimization of the descriptor
in respect of the number and the choice of views to be regis-
tered. In the following sect ions, we first describe the concept
of multiview 3D face descriptor, and then show how to op-
timize multiple single v iews to build 3D information using
our newly proposed “quasiview” concept, an extended term
of quasifrontal, which measures the influence power of a cer-
tain view to nearby views. Experimental results then follow.
(a) Five subregion
definitions on View
(0
◦
,0
◦
)
(b) Two subregion
definitions on View
(80
◦
,0
◦
)
Figure 3: Subregion definition depending on views superimposed
on the center face of our database.
2. MULTIVIEW 3D FACE DESCRIPTOR
The new descriptor we propose is called multiview 3D face
descriptor, which is supposed to have sufficient 3D informa-
tion of a face by describing the face as a mosaic of many
one-views as show n in Figure 1. This multiview 3D face de-
scriptor aims to cover any v iew between horizontal rotation

from
−90
◦
to 90
◦
and vertical rotation from −30
◦
to 30
◦
.
We note the range of such horizontal and vertical views as
[
−90
◦
···90
◦
]and[−30
◦
···30
◦
], respectively. The nota-
tion of [
·] is used to refer the range while (·)usedforaposi-
tion.
There are a few issues we encounter for the extension of
the conventional single view descriptor to multiview version.
(i) DB collect ion for training/test: there are not yet enough
data to be used for research on multiview face recogni-
tion. Most face database such as PIE, CMU, and YALE
has been built mainly for frontal views even though

nonfrontal face images are more usual.
(ii) Multiview face detector: to recognize a person from face
images, we first need to detect faces on photographs,
which is a rough alignment process.
(iii) View estimator: the view of the facial images should be
estimated.
(iv) Face alignment: faces are then aligned in predefined lo-
cation.
(v) Feature extraction: we extract features possibly de-
pending on views.
(vi) Descriptor optimization: we intend to produce efficient
descriptor containing views on horizontal rotation
[
−90
◦
···90
◦
] and vertical rotation [−30
◦
···30
◦
].
For DB generation, we could use 3D facial mesh models and
render them to get face images in arbitrary views. For the
experiment, the 3D facial mesh models of 108 subjects are
used and their rendered images are used for training and test
with 50/50 ratio. The database we use for the experiment is
described in our previous work [11, 12]aswellasposees-
timation and feature detection. In this paper, we focus on
the last two issues of feature extraction and descriptor op-

timization considering the various studies about multiview
face detections and view estimations. The most na
¨
ıve idea
to create multiview descriptor from a single-view one is the
W S. Lee and K A. Sohn 3
Normalized face
image f (x, y)
Subregion k parts
f
i
(x, y) of the image
Fourier transform
F(u, v)


F(u, v)


F
j
(u, v)


F
j
(u, v)


j = 1, , k

x
h
1
x
h
2
x
j
1
x
j
2
PCLDA
projection P
h
1
PCLDA
projection P
h
2
PCLDA
projection P
j
1
PCLDA
projection P
j
2
Vector normalization
y

h
1
y
h
2
y
j
1
y
j
2
LDA projection P
f
3
LDA projection P
j
3
Z
h
Z
i
Quantization Quantization
j
= 1, , k
Holistic Fourier feature jth subregion Fourier feature
Figure 4: Feature extraction used for multiview 3D face descriptor.
simple integration of N uniformly distributed single view de-
scriptors. If we register every 10
◦
apart, that is, if we use ev-

ery face image 10
◦
apart for our descriptor, we have to reg-
ister 19
× 7 views to cover the view space of horizontal ro-
tation [
−90
◦
···90
◦
] and vertical rotation [−30
◦
···30
◦
].
Then this very na
¨
ıve descriptor would result in size 133
× sin-
gle view descriptor size, which b ecomes too big to be used in
practice. Moreover, we could take advantage of the possibility
that some view regions might have larger coverage than oth-
ers so that we may need smaller number of views to describe
those regions. While the descriptor optimization is one of the
important steps for transition from single view to multiview
face descriptor, there has not been until now any published
result in this direction to the best of our knowledge. Here,
we aim to make use of our learning from frontal-view face
descriptors that a registered front view can be used to re-
trie ve nearby frontal views (quasifrontal) with high success

rate. Hence, we extend the concept of quasifrontal to qua-
siview and introduce some useful terms as follows.
(1) View mosaic. Mosaic of views 10
◦
apart covering hor-
izontal rotation [
−X
◦
···X
◦
] and vertical rotation
[
−Y
◦
···Y
◦
]. Here we choose X = 90 and Y = 30. It
can be visualized as shown in Figure 2.Thisviewmo-
saic is corresponding to any view (i.e., 3D) of a per-
son wherever the face is at least half. It is used later
on to check “quasiview” for each view in the view
mosaic.
(2) Quasiview with error rate K. It is an extension of
quasifrontal, from the frontal view to general views.
For instance, quasiview V
q
of a given (registered) view
V with error rate K means that faces on view V
q
can

be retrieved using a registered face in view V with ex-
pectederrorratelessthanorequaltoK. This will be
explored in Section 5
3. LOCALIZATION OF FACES IN MULTIVIEW
To use face images for training or as a query, we need to
extract and nor malize facial region. According to common
practice, positions of two eyes are used for normalization
such that the normalized image contains enough informa-
tion of the face but excludes unnecessary background. The
detailed localization specification is defined as follows.
(1) Size of images: 56
× 56.
(2) Positions of two eyes in the front view are on (0.3, 0.32)
and (0.7, 0.32) when width and height are considered
as 1.0. Here (, ) is used for (x, y) coordinates where the
numbers are between 0 and 1.
(3) Left eye position of the positive horizontal rotation
keeps (0.7, 0.32) while right eye position of the nega-
tive rotation does (0.3, 0.32).
(4) Vertical rotation has the same eye positions as the ones
on zero vertical rotation images.
Figure 2 summarizes the view mosaic of resulting local-
ized images for our view space of horizontal rotation
[
−90
◦
···90
◦
] and vertical rotation [−30
◦

···30
◦
].
4. FEATURE EXTRACTION
As an example of single view face descriptor, we use the
MPEG-7 advanced face recognition descriptor (AFR) [9]
which showed best performance in retrieval accuracy, speed
and data size as benchmarked by MPEG-7. More details can
befoundinMPEGdocument[9]. However, our focus in this
paper is to show how to build optimized integration of mul-
tiple views to recognize a face in any view based on single-
view face recognizers, so any single view face recognizer can
be used instead of MPEG-7 AFR.
4 EURASIP Journal on Image and Video Processing
(0.3, 0.32)
(0.7, 0.32)
0
20
40
60
80
100
120
0102030405060708090
h (degree)
Error rate 0.2
Error rate 0.05
(a) Horizontal rotation
(0.3, 0.32)
(0.7, 0.32)

0
20
40
60
80
100
120
−60 −40 −20 0 20 40 60
v (degree)
Error rate 0.2
Error rate 0.05
(b) Vertical rotation
Figure 5: Quasiview sizes with horizontal and vertical rotations. The x-axis in ( a) and (b) represents the degree of horizontal and vertical
rotation, respectively, and the y-axis shows the number of neighboring views which could be recognized by registering the view in x-axis
when allowed certain error rate (0.02 for blue plot, and 0.05 for red plot).
−30
◦
−20
◦
−10
◦
0
◦
10
◦
20
◦
30
◦
−90

◦
−80
◦
−60
◦
−40
◦
−20
◦
0
◦
20
◦
40
◦
60
◦
80
◦
90
◦
Trained view with holistic (5 features) + 5 subregions (5 features)
Trained view with holistic (5 features) + 5 subregions (2 features)
Trained view with holistic (5 features) + 2 subregions (5 features)
Registered
view
Figure 6: Views used for training and registr ation. 13 representative quasiviews are selected and used for training, and hence for registration.
The number of used features (especially, the features for subregions) also varies depending on the view.
For our experiment, MPEG-7 AFR is modified to adapt
to be multiview. AFR basically extrac ts features both in

Fourier space and luminance space. In the Fourier space,
features are extracted from the whole face, and luminance
spaceextractsfeaturesfromboththewholefaceandfive
subregions on the face as shown in Figure 3(a).Wesim-
plify, but also extend, this feature extraction algorithm to
our Subregion-based LDA on Fourier space for multiview pur-
pose. The biggest differences between the MPEG-7 AFR and
our model are (i) feature extraction in luminance space is
removed in our model; (ii) the subregion decomposition,
which was in luminance space, is now in Fourier space and
(iii) the number and positions of subregions are defined de-
pending on a given view, for example, for new frontal views,
we use the same five subregions as used in AFR, but for
near profile view, we only use two subregions as shown in
Figure 3(b). Figure 4 shows the overall feature extraction di-
agram. To summarize briefly, we first extract Fourier fea-
tures from both the whole face image and each subregion
of the image, and project all the features and their magni-
tudes using principle component—linear discriminant anal-
ysis (PCLDA) method. After normalizing the resulting vec-
tors, we do additional LDA projection, and finally quantize
them for descriptor efficiency. The first two modifications (i)
and (ii) give more efficient feature extraction method with
smaller descriptor size by extracting the same amount of in-
formation on a single space. The third modification (iii) is
caused by the multiview extension. If we use the same defini-
tion of subregion used in the front view for the profile view,
the background may seriously affect for recognition rate. So
we de fine different subregion depending on views as shown
in Figure 3.

W S. Lee and K A. Sohn 5
−80 −60 −40 −200 20406080
(0
◦
,0
◦
)
−40
−30
−20
−10
0
10
20
30
40
(a)
−80 −60 −40 −200 20406080
(60
◦
,0
◦
)
−40
−30
−20
−10
0
10
20

30
40
(b)
−80 −60 −40 −200 20406080
(30
◦
,30
◦
)
−40
−30
−20
−10
0
10
20
30
40
(c)
−80 −60 −40 −20 0 20406080
(30
◦
, −30
◦
)
−40
−30
−20
−10
0

10
20
30
40
(d)
−80 −60 −40 −200 20406080
(80
◦
,30
◦
)
−40
−30
−20
−10
0
10
20
30
40
(e)
−80 −60 −40 −200 20406080
(80
◦
, −30
◦
)
−40
−30
−20

−10
0
10
20
30
40
(f)
−80 −60 −40 −200 20406080
(80
◦
,0
◦
)
−40
−30
−20
−10
0
10
20
30
40
(g)
Figure 7: Representation of quasiviews. The x-axis and y-axis indicate the horizontal rotation from −90
◦
to 90
◦
and the vertical rotation
from
−40

◦
to 40
◦
, respectively. Big yellow spots represent the registered views and small red spots indicate corresponding quasiv iews with
error rate 0.05. The rectangles are the view region of interest in horizontal rotation [0
◦
···90
◦
] and vertical rotation [−30
◦
···30
◦
].
5. QUASIVIEW
Graham and Allinson [13] have calculated the distance be-
tween faces of different people over pose to predict the pose
dependency of a recognition system. Using the average Eu-
clidean distance between the people in the database over
the pose angles sampled, they predicted that faces should be
easiest to recognize around the 30
◦
range and consequently,
the best pose samples to use for an analysis should be concen-
trated around this range. Additionally, they expect that faces
are easier to recognize at the frontal view (0
◦
,0
◦
) than the
profile (90

◦
,0
◦
). Here, we use notation of (X
◦
, Y
◦
) to indicate
a view with X
◦
horizontal rotation and Y
◦
vertical rotation.
6 EURASIP Journal on Image and Video Processing
0 1020 30405060708090
Horizontal rotation
−30
−20
−10
0
10
20
30
Vert ical rot at ion
Figure 8: The region covered by 7 quasiviews in the view
mosaic of horizontal rotation [0
◦
···90
◦
] and vertical rotation

[
−30
◦
···30
◦
] with error rate 0.05. Registration with 7 views covers
93.93% of the view space, which means that we can retrieve faces
in any view represented in this plot from the registered descriptor
within allowed error rate 0.05.
Note that they have checked only horizontal rotation of hu-
man heads.
We use the new concept of “quasiview” corresponding to
the conventional “quasifrontal,” which is a measurement of
the influence of a registered view for recognition. To prove
that quasiview size depends on the view, we performed ex-
periments of quasiview inspection with accepted error rate
0.05, that is, we inspect the range of views that would be rec-
ognizable within error rate 0.05 g iven a view for registration.
Figure 5 shows how the quasiview size varies with pure hori-
zontal or vertical rotations of a head. To make fair compari-
son between different views, we extracted 24 holistic features
(without using subregion features) for each view. And im-
ages of nearby views are also included in the training of cer-
tain view (i.e., in obtaining the PCLDA basis for each view).
So for horizontal rotation, 9 views (the view of interest +8
nearby views) are used for training each view from (0
◦
,0
◦
)to

(70
◦
,0
◦
), 8 training views for the view (80
◦
,0
◦
), and 7 train-
ing views for the view (90
◦
,0
◦
). For vertical rotation, 9 tr ain-
ing views are used for each view from (0
◦
, −40
◦
)to(0
◦
,40
◦
),
8 training views for the views (0
◦
, −50
◦
)and(0
◦
,50

◦
), and 7
training views for the views (0
◦
, −60
◦
)and(0
◦
,60
◦
). Figure 5
is obtained before adding neighboring images in training.
Figure 6 can be helpful to understand which training views
are used for each registered view while it reflects our result
after optimization.
Figure 5 shows our quasiview measurements with syn-
thetically c reated (rendered) images of 108 3D facial mod-
els by rotating them into various angles. We counted the
number of nearby views which could be recognized when
we registered a certain view using two kinds of accepted
error rates 0.02 and 0.05. The result in Figure 5(a) shows
very similar pattern with the graph showing the average
distance between faces over view described in Graham and
Allinson’s paper [13]. The views (20
◦
,0
◦
) ∼ (30
◦
,0

◦
)have
both the biggest quasiview size and the biggest Euclidean
distance between the people in eigenspace among views
Figure 9: An example of registration. The views are needed in
the registration step to recognize a face in the 93.93% of the
view space where hor i zontal rotation [0
◦
···90
◦
], vertical rotation
[
−30
◦
···30
◦
], and their combined rotation of a head are allowed.
It means that we can retrieve a face in various poses within allowed
error rate 0.05 when we register only 7 views in a condition that a
given face is symmetric.
(0
◦
,0
◦
), (10
◦
,0
◦
), ,and(90
◦

,0
◦
). Figure 5(b) shows that
the views (0
◦
,0
◦
) ∼ (0
◦
,10
◦
) have the biggest quasiview size
among views (0
◦
, −60
◦
), (0
◦
, −60
◦
), ,and(0
◦
,60
◦
). The
views of heading downward have bigger quasiview size than
ones of heading upward and it makes us guess that it might
be easier to recognize people when they look downward more
than they look upward.
6. DESCRIPTOR OPTIMIZATION

Based on our study to check the quasiview size on horizontal
and vertical rotated heads, we now optimize the multiview
3D face descriptor by choosing several representative views
and recording the corresponding view specific features to-
gether. We have used the following selection criteria for reg-
istration views: we register views (i) with bigger quasiview
size for cost effect; (ii) which appear a lot in prac tice through
target environment analysis, for example, ATM, door access
control; (iii) considering efficient integra tion of quasiviews
covering the big region in view-mosaic; (iv) and which are
easy to register or easy to obtain. This choice is empirical
and we focus on covering the bigger range of face views with
more efficient face view registration. Remembering that our
features are extracted from PCLDA projections, we can select
the dimension of resulting features as we want. Hence, we can
also use variable feature numbers depending on the view. If
a view is easy to obtain for registration, but not so frequently
appear in practice, then we can use a smaller number of fea-
tures. More important views get bigger feature numbers.
In generating descriptors, training is considered as a step
to create space basis and matrix transform for feature extrac-
tion and as mentioned in Section 5 , many views are trained
for one registered view to increase the retrieval ability and
reliability. If we can embed more information in the step of
training, the registration can be done with smaller informa-
tion. For example, for the registered view (30
◦
,0
◦
), we use

9 surrounding views (10
◦
,0
◦
), (20
◦
,0
◦
), (30
◦
,0
◦
), (40
◦
,0
◦
),
(50
◦
,0
◦
), (30
◦
, −20
◦
), (30
◦
, −10
◦
), (30

◦
,10
◦
), (30
◦
,20
◦
)for
training. As summarized in Figure 6, for one view registra-
tion, the training is done with 6 to 9 views around the reg-
istered view. For this experiment, we have given three ways
to extract features based on basic feature extraction method
described in Section 4. Number of subregions and number
of features on subregions vary. So for some views, 5 holistic
features and 5 features for each of the five subregions are ex-
tracted which results in 30-dimensional view-specific feature
vector, and for other views 5, holistic features and 2 features
W S. Lee and K A. Sohn 7
for 5 subregions are extracted producing 15 dimensional vec-
tor. If a view is close to profile, we use 5 holistic features and
5 features for 2. For details for our experiment, see Figures 3
and 6. For one view, one image is selected.
In the experiment for multiview descriptor optimization
with rendered images from 3D facial models of 108 indi-
viduals, half of the images were used for training and the
other half were used for test. We show some examples of
quasiview in Figure 7 which shows the influence of each rep-
resented view. Big yellow spots are the views for registra-
tion and small red spots indicate corresponding quasiviews
with allowed error rate 0.05. Therefore, the region covered

by small spots surrounding a big spot indicates the influence
of the registered view (the big spot). For example, when we
register the very front view (the left most one in the middle
row in Figure 7), the horizontally 30
◦
-rotated and vertically
20
◦
-rotated views also could be recognized with error rate
0.05.
Through experiments with various combinations of qua-
siviews, a set of optimal views could be selected to create
final multiview 3D descriptor. An example of such possible
descriptor from the rendered images contains 13 views with
240-dimensional feature vector as shown in Figure 6.With
the allowed error rate 0.05, this descriptor was able to retrieve
the rendered images in the test database from 93.93% of the
viewsinviewmosaicofhorizontalrotation[
−90
◦
···90
◦
]
and vertical rotation [
−30
◦
···30
◦
]. Figure 8 shows the cov-
ered region of the views by the selected 7 views (right half of

the view space which corresponds to positive horizontal ro-
tation) considering the symmetry of the horizontal rotation.
Figure 9 shows an example which face views are needed for
registration to recognize the face in almost any pose. The 7
views are to be registered to recognize a face in the 93.93% of
the view space where horizontal rotation [0
◦
···90
◦
], verti-
cal rotation [
−30
◦
···30
◦
], and their combined rotation of
a head are allowed. It means that we can retrieve a face in
various poses within allowed error rate 0.05 when we reg-
ister only 7 views in a condition that a given face is sym-
metric. For a reference, when we allowed error rate of 0.1, it
covers 95.36% of the view space, 97.57% for error rate 0.15,
and 97.98% for error rate 0.2. For the experiment, the test-
ing views are situated at intervals 5 degrees while a 10-degree
interval is used for training.
As a reference, the MPEG-7 AFR [9, 10] has 48 dimen-
sions with error rate 0.3013 and 128 dimensions with er-
ror rate 0.2491 for photograph images. Here we used the er-
ror rate of ANMRR (average normalized modified retrieval
rank), the MPEG-7 retrieval metric, which indicates how
manyofthecorrectimagesareretrievedaswellashowhighly

they are ranked among the retrieved ones. Details about AN-
MRR can be found in MPEG related documents like [14].
7. CONCLUSION
We have shown how the single-view face descriptor could be
extended to multiview one in efficient way by checking the
size of quasiview, which is a measure of the view influence.
For the experiment, the 3D facial mesh models of 108 sub-
jects are used and their rendered images are used for training
and test with 50/50 ratio. Only 13 views could be chosen as
registered views throughout our optimization. This descrip-
tor in 240 dimensions is able to retrieve images of 93.93%
views of total region of view mosaic of horizontal rotation
from
−90
◦
to 90
◦
and vertical rotation from −30
◦
to 30
◦
within error rate 0.05.
The aim of this new descriptor is to be used to extract a
face in any view by containing compact 3D information by
optimization for how many and which views are to be regis-
tered. The extension to multiview is not very costly in terms
of number of registration views thanks to the quasiview anal-
ysis. Even though we have used a specific face descriptor for
the experiment, the potential of this method enables us to
include any available 2D face recognition methods by show-

ing how to combine them in optimized way by checking qua-
siview size. Ongoing research includes new feature extraction
methods for profile views and missing view interpolation in
the registration step.
REFERENCES
[1] A. Samal and P. A. Iyengar, “Automatic recognition and anal-
ysis of human faces and facial expressions: a survey,” Pattern
Recognition, vol. 25, no. 1, pp. 65–77, 1992.
[2] S. Z. Li, L. Zhu, Z. Q. Zhang, A. Blake, H. J. Zhang, and H.
Shum, “Statistical learning of multi-view face detection,” in
Proceedings of the 7th European Conference on Computer Vision
(ECCV ’02), vol. 4, pp. 67–81, Copenhagen, Denmark, May
2002.
[3] Y. Li, S. Gong, and H. Liddell, “Support vector regression and
classification based multi-view facedetection and recognition,”
in Proceedings of the 4th IEEE International Conference on Au-
tomatic Face and Gesture Recognition, pp. 300–305, Grenoble,
France, March 2000.
[4] G. Shakhnarovich, L. Lee, and T. Darrell, “Integrated face and
gait recognition from multiple views,” in Proceedings of the
IEEE Computer Society Conference on Computer Vision and
Pattern Recognition (CVPR ’01), vol. 1, pp. 439–446, Kauai,
Hawaii, USA, December 2001.
[5] V. Blanz and T. Vetter, “Face recognition based on fitting a 3D
morphable model,” IEEE Transactions on Pattern Analysis and
Machine Intelligence, vol. 25, no. 9, pp. 1063–1074, 2003.
[6] A. M. Bronstein, M. M. Bronstein, and R. Kimmel,
“Expression-invariant 3D face recognition,” in Proceedings of
the 4th International Conference on Audio- and Video-Based
Biometric Person Authentication (AVBPA ’03), vol. 2688 of Lec-

ture Notes in Computer Science, pp. 62–69, Guildford, UK, June
2003.
[7]D.M.GavrilaandL.S.Davis,“3-Dmodel-basedtracking
of humans in action: a multi-view approach,” in Proceedings
of the IEEE Computer Society Conference on Computer Vision
and Pattern Recognition (CVPR ’96), pp. 73–80, San Francisco,
Calif, USA, June 1996.
[8] K. W. Bowyer, K. Chang, and P. Flynn, “A survey of approaches
and challenges in 3D and multi-modal 3D + 2D face recog-
nition,” Computer Vision and Image Understanding, vol. 101,
no. 1, pp. 1–15, 2006.
[9] A. Yamada and L. Cieplinski, “MPEG-7 Visual part of eXper-
imentation Model Version 17.1,” ISO/IEC JTC1/SC29/WG11
M9502, Pattaya, Thailand, March 2003.
[10] T. Kamei, A. Yamada, H. Kim, W. Hwang, T K. Kim, and S. C.
Kee, “CE report on Advanced Face Recognition Descriptor,”
8 EURASIP Journal on Image and Video Processing
ISO/IEC JTC1/SC29/WG11 M9178, Awaji, Japan, December
2002.
[11] W S. Lee and K A. Sohn, “Face recognition using computer-
generated database,” in Proceedings of Computer Graphics In-
ternational (CGI ’04), pp. 561–568, IEEE Computer Society
Press, Crete, Greece, June 2004.
[12] W S. Lee and K A. Sohn, “Database construction & recogni-
tion for multi-view face,” in Proceedings of the 6th IEEE Inter-
national Conference on Automatic Face and Gesture Recognition
(FGR ’04), pp. 350–355, IEEE Computer Society Press, Seoul,
Korea, May 2004.
[13] D. B. Graham and N. M. Allinson, “Characterizing virtual
eigensignatures for general purpose face recognition,” in Face

Recognition: From Theory to Applications,H.Wechsler,P.J.
Phillips, V. Bruce, F. Fogelman-Soulie, and T. S. Huang, Eds.,
pp. 446–456, Springer, Berlin, Germany, 1998.
[14] G. Park, Y. B aek, and H K. Lee, “A ranking algorithm using
dynamic clustering for content-based image retrieval,” in Pro-
ceedings of the International Conference Image and Video Re-
trieval (CIVR ’02), vol. 2383 of Lecture Notes in Computer Sci-
ence, pp. 328–337, London, UK, July 2002.